Transformer unbalanced operation winding vibration form identification method based on Autoformer generalization model

Through the Autoformer generalization model combined with the information of electromagnetic and mechanical subsystems, the vibration forms of the transformer winding are identified, which solves the problem of difficult to identify the vibration forms under the unbalanced operating state of the winding in the prior art, and achieves a high-accuracy recognition effect.

CN119089786APending Publication Date: 2024-12-06NORTHEAST DIANLI UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411190598.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-28
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The prior art is difficult to effectively identify the vibrational form of the power transformer winding in an unbalanced operating state, resulting in difficulty in identifying faults.

Method used

Using the Autoformer generalization model, by dividing the transformer into electromagnetic and mechanical subsystems, a digital simulation and dynamic mode experimental platform is built, multi-source physical information is obtained, electrical and vibration signals are extracted as key characteristic parameters, and fusion characteristic indicators are constructed, based on the Autoformer model Clustering and dividing the sample space to realize the identification of winding vibration patterns.

Benefits of technology

The accuracy of the vibration pattern recognition of the transformer windings has been improved to reach 97.19%, effectively solving the problem of difficult vibration pattern recognition in the unbalanced operating state of the windings.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089786A_ABST
    Figure CN119089786A_ABST
Patent Text Reader

Abstract

A transformer unbalanced operation winding vibration form recognition method based on an Autoformer generalization model belongs to the technical field of transformers, and evaluates the vibration form of a winding by extracting electromagnetic-mechanical information features of a transformer, forming a global mode-state feature information library, fusing multivariate feature information and combining the Autoformer generalization model. And a form identification scheme based on information physical fusion is formed. According to the scheme, the winding vibration form can be effectively and accurately identified.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention belongs to the technical field of transformers, and in particular relates to a method for identifying vibration forms of transformer unbalanced operation windings based on an Autoformer generalized model. Background Art

[0002] Power transformers are the core equipment of power grids, and their operating status affects the safe and stable operation of the system. Existing statistics and accident analysis show that mechanical failure of windings is an important factor leading to transformer failures. Therefore, the identification and evaluation of transformer winding vibration morphology has important research value and engineering significance. Using observable electrical parameters to describe difficult-to-observe physical characteristics is currently a key issue in transformer winding operating status monitoring and identification. Summary of the invention

[0003] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art, provide a scientific, reasonable, efficient and practical method for solving the technical problem that the vibration morphology of the winding of the power transformer in the unbalanced operation state is difficult to identify in the prior art, and provide a transformer unbalanced operation winding vibration morphology identification method based on the Autoformer generalization model, divide the transformer virtual / physical entity into electromagnetic and mechanical subsystems, and obtain multi-source and multi-feature physical information parameters by building a digital simulation and dynamic model experimental platform to form a global modal-state information library. The electrical and vibration signals are extracted as the key feature parameters of the winding information fusion, and the corresponding morphology of the electromagnetic-mechanical sample space is quantitatively evaluated by the fusion feature index. Based on an Autoformer generalization model, the morphological domain mapped by each sample space is clustered and divided to form a transformer three-phase unbalanced operation vibration morphology identification scheme. The morphological domain mapped by each sample space is clustered and divided to form a vibration morphology identification scheme based on the Autoformer generalization model, thereby improving the recognition accuracy.

[0004] A transformer unbalanced operation winding vibration morphology recognition method based on an Autoformer generalized model comprises the following steps, and the following steps are performed in sequence:

[0005] Step 1: Electromagnetic-mechanical information extraction and consistency verification

[0006] The winding is divided into electromagnetic and mechanical subsystems, and a virtual simulation model is established to obtain the multi-source characteristic information of the transformer. The electromagnetic information includes the port voltage u, current i, load rate η, unbalance λ and leakage flux B σ , the mechanical information includes the winding electromagnetic force F and the vibration acceleration g. Information fusion is used to check the spatiotemporal consistency of virtual and physical entities and multi-source feature information, verify the validity of the acquisition of virtual and physical entities and feature information, and then build a global modal-state information library with time as the index;

[0007] Step 2: Information Fusion

[0008] The moving average function of the progressive decomposition module in the Autoformer generalized model is used to decompose the collected information in the time domain to reflect its periodic evolution law and cumulative development trend.

[0009] Preprocessing: Use λ as the label to search the modal-state information library, extract the current information as the feature vector X and the vibration information as the feature vector Y, form the sample set Z, and normalize it to get z n , and then calculate the fusion characteristic index ξ, set the transformer winding vibration form corresponding to different ξ value ranges, and the calculation formula of ξ is:

[0010]

[0011] Where h is the fusion weight; z n is the nth sample;

[0012] Encoder / decoder: The sample set is passed through the encoder to obtain the periodic evolution characteristics of the original sequence. An adaptive learning mechanism is introduced based on the association attributes of i-λ and ig to form a modal-state information feature mapping agent. The parallel output information is used to assist the decoder in making decisions.

[0013] Classifier: Based on the fusion feature index ξ, the self-correction and generalization representation modules are embedded. The self-correction results are compared with the decoder output to improve the accuracy of morphological recognition. The generalization representation is fused with the morphological recognition results to improve the generalization ability of class domain division.

[0014] Step 3: Winding Vibration Identification

[0015] By extracting the electromagnetic-mechanical information features of the transformer, a global modal-state feature information database is formed. The multivariate feature information is integrated, and the winding vibration morphology is evaluated using the Autoformer generalization model to form a morphology recognition scheme based on information-physics fusion.

[0016] Build a physical entity dynamic model platform to measure the electrical information and vibration information of the transformer winding; build a virtual entity based on the physical entity, solve the magnetic field model to obtain the winding leakage magnetic field and electromagnetic force, calculate the mechanical model through electromagnetic-mechanical sequential coupling to obtain vibration information, and form a global winding mode-state information library based on labels and indexes;

[0017] The characteristics of the electromagnetic subsystem are characterized by the winding current, and the vibration characteristics of the mechanical subsystem are characterized by the vibration acceleration. Combined with the virtual and real interactive information, the fusion feature index is calculated and the mapping relationship is constructed to form a fusion feature information database of the winding unbalanced operation.

[0018] The time domain characteristics of winding current signals and vibration signals are analyzed, and the Autoformer generalized recognition model is trained by combining virtual and physical sample sets to realize the recognition and prediction of transformer winding vibration morphology.

[0019] Furthermore, the specific process performed by the classifier is:

[0020]

[0021] Where Avgpool is the average function and Padding is the fixed vector length function. t is the cumulative component, z s is the periodic component; z is the sample, and equation (4) is recorded as the sequence decomposition: (z t , z s )=SeriesDe(z), where SeriesDe is the sequence decomposition and is embedded between Autoformer layers.

[0022]

[0023]

[0024]

[0025] In the formula, Correlation is the autocorrelation operation, Feed is the feedforward module, and Concat represents the connection function; It is the periodic regular output part of the lth layer obtained after a sequence decomposition module eliminates the cumulative trend part; Input for the l-1th layer encoder; is the periodic regular output part of the cumulative trend elimination part obtained after two sequence decomposition modules in the lth layer; “_” is the cumulative trend elimination part; z en is the encoder input, P en is the periodic output part of the encoder; z ens is the periodic regular part of the sample, z ent is the cumulative trend part; z0 is the zero value of the encoder decomposition result; the zero value and mean value of the encoder decomposition result correspond to z0 and z mean ; It is the periodic regular output part obtained after passing through a sequence decomposition module in the lth layer; It is the cumulative trend information output part obtained after a sequence decomposition module in the lth layer; Input the periodic regularity feature information of the l-1 layer decoder; It is the periodic regular output part obtained after two sequence decomposition modules in the lth layer; It is the cumulative trend information output part obtained after two sequence decomposition modules in the lth layer; Input the periodic regular feature information of the Nth layer decoder; It is the periodic characteristic component finally output by the decoder; is the cumulative trend feature component of the final output of the decoder; de Input the periodic regularity characteristic information of the decoder; It is the output part of the cumulative trend information obtained in the lth layer; is the cumulative trend information output part obtained in the l-1th layer; W l,1 is the weight obtained after passing through a sequence decomposition module in the lth layer; W l,2 is the weight obtained after two sequence decomposition modules in the lth layer; W l,3 is the weight obtained after three sequence decomposition modules in the lth layer; T de is the cumulative trend information input; l represents the number of decoder layers, and W is the weight matrix.

[0026] Furthermore, the auxiliary decoder decision includes: the upper layer operation gradually extracts and accumulates the cumulative development trend information from the samples, the lower layer uses the autocorrelation operation to extract the higher-order time dependency of the encoder output, and finally outputs it through the feedforward network, considering the cumulative trend development of the transformer components under the given scenario based on the periodic regular changes of the observed information.

[0027] Through the above design scheme, the present invention can bring the following beneficial effects: the multi-physical feature fusion parameter method based on the winding current-mapping vibration signal combined with the Autoformer generalized recognition model can effectively identify the winding vibration form, and the accuracy can be improved to 97.19%. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 It is a schematic diagram of electromagnetic-mechanical information-physics fusion of a transformer unbalanced operation winding vibration morphology recognition method based on the Autoformer generalized model of the present invention.

[0029] Figure 2 The invention discloses a transformer finite element virtual entity diagram of a transformer unbalanced operation winding vibration morphology identification method based on an Autoformer generalized model.

[0030] Figure 3 The present invention provides a schematic diagram of a global modal information library for a transformer unbalanced operation winding vibration morphology recognition method based on an Autoformer generalized model.

[0031] Figure 4The present invention discloses an i-λ-F information association diagram of a transformer unbalanced operation winding vibration morphology recognition method based on an Autoformer generalized model.

[0032] Figure 5 The present invention discloses an i-λ-g information association diagram of a transformer unbalanced operation winding vibration morphology recognition method based on an Autoformer generalized model.

[0033] Figure 6 The present invention discloses a primary winding fusion characteristic index diagram of a transformer unbalanced operation winding vibration morphology recognition method based on an Autoformer generalized model.

[0034] Figure 7 The present invention discloses a secondary winding fusion characteristic index diagram of a transformer unbalanced operation winding vibration morphology recognition method based on an Autoformer generalized model.

[0035] Figure 8 The present invention is an Autoformer generalized model diagram of a transformer unbalanced operation winding vibration morphology recognition method based on the Autoformer generalized model.

[0036] Fig. 9 This is a winding vibration morphology recognition scheme diagram of a transformer unbalanced operation winding vibration morphology recognition method based on an Autoformer generalized model of the present invention. DETAILED DESCRIPTION

[0037] In order to better understand the purpose, structure and function of the present invention, the following Figure 1-9 The present invention further illustrates a method for identifying the vibration morphology of a transformer unbalanced operation winding based on an Autoformer generalized model.

[0038] Transformer electromagnetic-mechanical information-physical fusion includes electromagnetic and mechanical physical information and virtual (virtual simulation) and real (physical monitoring) data fusion. The basic principles of transformer information-physical fusion are as follows: Figure 1 As shown in the figure, the winding is divided into an electromagnetic subsystem and a mechanical subsystem, and key physical characteristic information is extracted respectively, wherein the electromagnetic information includes the port voltage u, current i, load rate η, unbalance λ and leakage flux Bσ; the mechanical information includes the winding electromagnetic force F and vibration acceleration g. In the process of virtual and physical data extraction, the physical entity can directly measure the characteristic information of u, i, λ, η and g, and the physical entity includes the dynamic model test platform and / or test transformer of the transformer; the virtual entity obtains the physical entity characteristic information u, i, λ, η and g through simulation, and the virtual entity includes the simulated transformer, and obtains Bσ through simulation. σand F. Information fusion is used to check the spatiotemporal consistency of virtual and physical entities and multi-source feature information, which includes feature information of physical entities and simulated entities, to verify the validity of virtual and physical entities and feature information acquisition, and then a global modal-state information database is constructed to lay the foundation for cyber-physical fusion.

[0039] 1. Electromagnetic-mechanical feature information extraction

[0040] 1.1 Imbalance evaluation method

[0041] Taking the three-phase unbalanced operation mode of the transformer as the research object, the evaluation method of the unbalance degree λ is as follows:

[0042]

[0043] Where I max is the maximum current, I min is the minimum current. When the unbalanced phase current exceeds the normal operating phase current, λ is positive, otherwise it is negative. In order to improve the accuracy of the comparison between virtual simulation and physical experiment, a closed-loop solution is adopted: modeling based on actual conditions → simulation verification based on actual measurements → calibration of virtual entities → agreement between virtual and real entities → feedback prediction of actual conditions.

[0044] 1.2 Physical Entity Information Acquisition

[0045] Take the three-phase three-column experimental transformer (JSSG-15kVA 380V / 110V) as an example, in which the positive Young's modulus and Poisson's ratio are obtained through material tension and compression tests.

[0046] Table 1 Three-phase transformer parameters

[0047]

[0048] The information collection process is as follows:

[0049] 1) Set different operation modes of the transformer, debug the sensor and current monitoring module (ZDS2022), the technical index parameters of ZDS2022 are shown in Table 2, collect port voltage and current information in real time, filter the voltage and current, and remove noise and interference information;

[0050] Table 2 ZDS2022 technical parameters

[0051]

[0052]

[0053] 2) Combined with the unbalanced operation of the transformer, the piezoelectric acceleration sensor is connected and the vibration monitoring module (ZT902) is debugged. The technical index parameters of ZT902 are shown in Table 3. The transformer winding vibration information is collected in real time under different modes, and the vibration signal is processed for noise reduction;

[0054] 3) Using the measured transformer component material properties, as shown in Table 1;

[0055] 4) Using the imbalance λ as a label and time t as an index, the electromagnetic-mechanical information is stored.

[0056] Table 3 ZT902 technical parameters

[0057]

[0058] 1.3 Virtual Entity Information Acquisition

[0059] Digital modeling is carried out by actually measuring the electrical, mechanical and material data of the transformer. Through a large number of simulation and experimental comparisons, the model parameters and field settings are repeatedly revised to achieve the consistency of the simulation in terms of entity and operating characteristics, so that the simulation model can effectively reflect the electromagnetic mechanical process of the physical entity.

[0060] The finite element simulation software Comsol is used to construct a proportional finite element virtual entity, such as Figure 2 The physical parameters of the transformer (silicon steel sheet model is 50WW800) are shown in Table 1. The virtual entity is divided into a magnetic field connection domain, a current excitation domain, and a mechanical stress domain, corresponding to the core, winding and other components, and the space is consistent with the physical entity.

[0061] The virtual entity material parameter settings are consistent with the information in Table 1. During the simulation process, the mesh division uses differentiated tetrahedral meshes. The core winding components that need to be accurately calculated select "fine" units, and the pads and pull plates select "rough" units. The perfect matching layer uses hexahedral sweeping. The solver selects the "full coupling solution method". Apply excitation and boundary: 1) The winding excitation in the electromagnetic system is set to circular current, and the magnetic field boundary is set to the parallel boundary condition of the magnetic field lines, and the rest are set to natural boundary conditions; 2) The transformer base in the mechanical system is set to a fixed constraint condition, and the rest is subject to the roller support constraint condition, that is, the transformer base is fixed, and only the axial vibration of the component is considered.

[0062] 1.4 Building a global modal information library

[0063] The transformer mechanical subsystem is solved by electromagnetic-mechanical sequential coupling. Figure 3 As shown, the specific steps are as follows:

[0064] 1) Using the imbalance degree λ as the control parameter and state label, and time t as the index, the electromagnetic sub-information library is based on t k The winding current is used as the excitation to set the current-carrying excitation domain and the magnetic field connection domain, solve the spatiotemporal distribution of the winding magnetic field and obtain the electromagnetic field characteristic parameters, such as the current i k 、Magnetic leakage B σk wait. I 0为 Initial current, u(t) is the port voltage at time t.

[0065] 2) Information acquisition iteration process. If the absolute convergence norm is less than the convergence criterion value, or the coupling cycle reaches the set number of times, the iteration ends. A link is established based on the time point index, and the calculation results are stored in the global mode-state information library. Otherwise, the next time index t is updated. k+1 , the current i k+1 The magnetic field model is input as coupling parameters to solve the magnetic field at the next moment.

[0066] 3) Electromagnetic-mechanical coupling simulation. Retrieve t in the global modal information library k The electromagnetic information at the moment is used to solve the temporal and spatial distribution of the electromagnetic force of the winding and to calculate the electromagnetic force at t k It is a dynamic storage for tags. k The electromagnetic force at the moment is used as the excitation to set the mechanical stress domain, solve the winding vibration distribution and obtain the mechanical field characteristic parameter vibration acceleration g. If the absolute convergence norm is less than the convergence criterion value, or the coupling cycle reaches the set number of times, the iteration ends. Otherwise, the next time index t is updated. k+1 , enter the next iterative calculation.

[0067] 4) Using λ as label and t as index, the electromagnetic and mechanical feature information are combined to form a global modal information library.

[0068] 2 Morphological Recognition Based on Autoformer Generalization Model

[0069] 2.1 Construction of sample set

[0070] Use λ as the label to search the modal-state information library, extract the current information as the feature vector X and the vibration information as the feature vector Y, form the sample set Z, and normalize it to get the nth sample z n .

[0071]

[0072] In the formula, z i is the i-th sample, z min Sample minimum value, z max The maximum value of the sample.

[0073] 2.2 Multi-physics association and fusion characteristic indicators

[0074] Based on the global modal-state information, the mapping relationship between winding current and leakage magnetic field and electromagnetic force is constructed respectively, and the changes of electromagnetic and mechanical subsystems are reflected by winding current and vibration respectively. Figure 4 As shown. Comprehensive electromagnetic subsystem information (u, i, B σ ) and mechanical subsystem information (F, g) to form a global modal information library, extract key response parameters current i and vibration acceleration g, and construct a fusion characteristic index ξ, as shown in formula (3) and the attached Figure 5 h is the fusion weight. According to ξ, there are four types of winding vibration forms, namely normal operation, risk warning, abnormal warning and fault alarm. The abnormal warning includes abnormal warning A and abnormal warning B, and the risk warning includes risk warning A and risk warning B, as shown in Table 4.

[0075] Table 4 Table of ξ values ​​for different winding vibration forms

[0076]

[0077]

[0078] In the formula, h i is the fusion weight of the i-th sample.

[0079] 2.3 Morphological Recognition of Autoformer Generalization Model

[0080] In order to improve the recognition generalization and more accurately reflect the internal operating conditions of the transformer, the Autoformer generalized model is used to identify the winding vibration form based on the global mode-state fusion information data, multi-physics association relationship and fusion feature index. Figure 6 According to the characteristics of winding vibration morphology during the unbalanced operation of the transformer, the moving average function of the progressive decomposition module in the Autoformer generalized model is used to decompose the collected information in the time domain to reflect its periodic evolution law and cumulative development trend.

[0081] Encoder / decoder: The sample set obtains the periodic evolution characteristics of the original sequence through the encoder in the Autoformer model. The original sequence is the waveform of the sample input at the beginning. Based on the association attribute of i-λ and ig, an adaptive learning mechanism is introduced to form a modal-state information feature mapping agent, that is, through the association attribute of i-λ and ig, a fitting function is formed. The fitting function is used to assist the encoder in discrimination, and the parallel output information is used to assist the decoder in decision-making: 1) The upper layer operation mainly extracts and accumulates the cumulative development trend information from the sample step by step; 2) The lower layer uses the autocorrelation operation to extract the higher-order time dependency of the encoder output, and finally outputs it through the feedforward network. Based on the periodic regular changes of the observed information, the cumulative trend development of the transformer components under the given scenario is considered.

[0082] Classifier: Based on the fusion feature index ξ, the self-correction and generalization representation modules are embedded. The self-correction result is compared with the decoder output to improve the accuracy of morphological recognition. The generalization representation is fused with the morphological recognition result to improve the generalization ability of class domain division. The specific process is shown in formulas (4)-(7).

[0083]

[0084] Where Avgpool is the average function and Padding is the fixed vector length function. t is the cumulative component, z s is the periodic component, and equation (4) is recorded as the sequence decomposition: (z t , z s )=SeriesDe(z), where SeriesDe is the sequence decomposition and is embedded between Autoformer layers.

[0085]

[0086]

[0087]

[0088] In the formula, Correlation is the autocorrelation operation, Feed is the feedforward module, and Concat represents the connection function; It is the periodic regular output part of the lth layer obtained after a sequence decomposition module eliminates the cumulative trend part; Input for the l-1th layer encoder; is the periodic regular output part of the cumulative trend elimination part obtained after two sequence decomposition modules in the lth layer; “_” is the cumulative trend elimination part; z en is the encoder input, P en is the periodic output part of the encoder; z ens is the periodic regular part of the sample, z ent is the cumulative trend part; z0 is the zero value of the encoder decomposition result; the zero value and mean value of the encoder decomposition result correspond to z0 and z mean ; It is the periodic regular output part obtained after passing through a sequence decomposition module in the lth layer; It is the cumulative trend information output part obtained after a sequence decomposition module in the lth layer; Input the periodic regularity feature information of the l-1 layer decoder; It is the periodic regular output part obtained after two sequence decomposition modules in the lth layer; It is the cumulative trend information output part obtained after two sequence decomposition modules in the lth layer; Input the periodic regular feature information of the Nth layer decoder; It is the periodic characteristic component finally output by the decoder; is the cumulative trend feature component of the final output of the decoder; de Input the periodic regularity characteristic information of the decoder; It is the output part of the cumulative trend information obtained in the lth layer; is the cumulative trend information output part obtained in the l-1th layer; W l,1 is the weight obtained after passing through a sequence decomposition module in the lth layer; W l,2 is the weight obtained after two sequence decomposition modules in the lth layer; W l,3 is the weight obtained after three sequence decomposition modules in the lth layer; T de Input for cumulative trend information; P l,3de and T lde They are the periodic feature component and cumulative trend feature component of the final output of the decoder respectively; l represents the number of decoder layers, and W is the weight matrix.

[0089] 2.4 Winding vibration morphology identification scheme

[0090] By extracting the electromagnetic-mechanical information features of the transformer, a global modal-state feature information library is formed, and multi-dimensional feature information is integrated to form a morphological recognition solution based on the Autoformer generalization model. Figure 7 shown.

[0091] 1) Build a physical entity dynamic model platform to measure the electrical information and vibration information of the transformer winding; build a virtual entity based on the physical entity, solve the magnetic field model to obtain the winding leakage flux and electromagnetic force, calculate the mechanical model through electromagnetic-mechanical sequential coupling to obtain vibration information, and form a global mode-state information library of the winding according to labels and indexes.

[0092] 2) The characteristics of the electromagnetic subsystem are characterized by the winding current, and the vibration characteristics of the mechanical subsystem are characterized by the vibration acceleration. Combined with the virtual and real interactive information, the fusion feature index is calculated and the mapping relationship is constructed to form a fusion feature information library of the winding unbalanced operation.

[0093] 3) Analyze the time domain characteristics of the winding current signal and vibration signal, combine the virtual / real sample sets, train the Autoformer generalization recognition model, and realize the recognition and prediction of the transformer winding vibration morphology.

[0094] The calculation conditions, legends, etc. in the embodiments of the present invention are only used to further illustrate the present invention, but are not exhaustive and do not constitute a limitation on the scope of protection of the claims. Those skilled in the art can think of other substantially equivalent alternatives based on the inspiration obtained from the examples of the present invention without creative work, and all of them are within the scope of protection of the present invention.

Claims

1. A transformer unbalanced operation winding vibration morphology recognition method based on the Autoformer generalization model, characterized by: The method comprises the following steps, and the following steps are performed in sequence: Step 1: Electromagnetic-mechanical information extraction and consistency verification The winding is divided into electromagnetic and mechanical subsystems, and a virtual simulation model is established to obtain the multi-source characteristic information of the transformer. The electromagnetic information includes the port voltage u, current i, load rate η, unbalance λ and leakage flux B σ , the mechanical information includes the winding electromagnetic force F and the vibration acceleration g. Information fusion is used to check the spatiotemporal consistency of virtual and physical entities and multi-source feature information, verify the validity of the acquisition of virtual and physical entities and feature information, and then build a global modal-state information library with time as the index; Step 2: Information Fusion The moving average function of the progressive decomposition module in the Autoformer generalization model is used to decompose the collected information in the time domain; Preprocessing: Use λ as the label to search the modal-state information library, extract the current information as the feature vector X and the vibration information as the feature vector Y, form the sample set Z, and normalize it to get z n , and then calculate the fusion characteristic index ξ, set the transformer winding vibration form corresponding to different ξ value ranges, and the calculation formula of ξ is: Where h is the fusion weight; z n is the nth sample; Encoder / decoder: The sample set is passed through the encoder to obtain the periodic evolution characteristics of the original sequence. An adaptive learning mechanism is introduced based on the association attributes of i-λ and ig to form a modal-state information feature mapping agent. The parallel output information is used to assist the decoder in making decisions. Classifier: Based on the fusion feature index ξ, the self-correction and generalization representation modules are embedded. The self-correction results are compared with the decoder output to improve the accuracy of morphological recognition. The generalization representation is fused with the morphological recognition results to improve the generalization ability of class domain division. Step 3: Winding Vibration Identification Build a physical entity dynamic model platform to measure the electrical information and vibration information of the transformer winding; build a virtual entity based on the physical entity, solve the magnetic field model to obtain the winding leakage magnetic field and electromagnetic force, calculate the mechanical model through electromagnetic-mechanical sequential coupling to obtain vibration information, and form a global winding mode-state information library based on labels and indexes; The characteristics of the electromagnetic subsystem are characterized by the winding current, and the vibration characteristics of the mechanical subsystem are characterized by the vibration acceleration. Combined with the virtual and real interactive information, the fusion feature index is calculated and the mapping relationship is constructed to form a fusion feature information database of the winding unbalanced operation. The time domain characteristics of winding current signals and vibration signals are analyzed, and the Autoformer generalized recognition model is trained by combining virtual and physical sample sets to realize the recognition and prediction of transformer winding vibration morphology.

2. According to claim 1, a transformer unbalanced operation winding vibration morphology recognition method based on the Autoformer generalization model is characterized by: The specific process performed by the classifier is: Where Avgpool is the average function, Padding is the fixed vector length function; t is the cumulative component; z s is the periodic component; z is the sample, and equation (4) is recorded as the sequence decomposition: (z t , z s ) = SeriesDe(z), where SeriesDe is the sequence decomposition and is embedded between Autoformer layers; In the formula, Correlation is the autocorrelation operation, Feed is the feedforward module, and Concat represents the connection function; It is the periodic regular output part of the lth layer obtained after a sequence decomposition module eliminates the cumulative trend part; Input for the l-1th layer encoder; is the periodic regular output part of the first layer after two sequence decomposition modules to eliminate the cumulative trend part; P en It is the periodic output part of the encoder; ens is the periodic regular part of the sample, z ent is the cumulative trend part; z0 is the zero value of the encoder decomposition result; the zero value and mean value of the encoder decomposition result correspond to z0 and z mean ; It is the periodic regular output part obtained after passing through a sequence decomposition module in the lth layer; It is the cumulative trend information output part obtained after a sequence decomposition module in the lth layer; Input the periodic regularity feature information of the l-1 layer decoder; It is the periodic regular output part obtained after two sequence decomposition modules in the lth layer; It is the cumulative trend information output part obtained after two sequence decomposition modules in the lth layer; Input the periodic regular feature information of the Nth layer decoder; It is the periodic characteristic component finally output by the decoder; It is the cumulative trend feature component of the final output of the decoder; z de Input the periodic regularity characteristic information of the decoder; It is the output part of the cumulative trend information obtained in the lth layer; is the cumulative trend information output part obtained in the l-1th layer; W l,1 is the weight obtained after passing through a sequence decomposition module in the lth layer; W l,2 is the weight obtained after two sequence decomposition modules in the lth layer; W l,3 is the weight obtained after three sequence decomposition modules in the lth layer; T de Input for cumulative trend information.

3. According to claim 1, a transformer unbalanced operation winding vibration morphology recognition method based on the Autoformer generalization model is characterized by: The auxiliary decoder decision includes: the upper layer operation gradually extracts and accumulates the cumulative development trend information from the samples, and the lower layer uses the autocorrelation operation to extract the higher-order time dependency of the encoder output, and finally outputs it through the feedforward network.